[gmx-users] Re:200 CPU, 3ns/day for 80,000 atoms

[Dr. Vitaly V. Chaban]

>  I am using gromacs for membrane simulation (under CHARMM36 FF) which
> contains around 80,000 atoms. I've submitted over 200 CPU in the cluster
> for such system with 2 fs time step. And what really astonished is that the
> efficiency for such simulation is only 3ns/day..... I am wondering what
> happen to my system or gromacs? What can I do to fasten the simulation?
>
> here is my md.mdp:
> *
> title                    = god!
> cpp                      = /usr/bin/cpp
> include                  =
> define                   =
> integrator               = md
> dt                       = 0.001
> nsteps                   = 100000000
> nstxout                  = 1000000
> nstvout                  = 1000000
> nstlog                   = 1000000
> nstenergy                = 10000
> nstxtcout                = 100000
> xtc_grps                 =
> energygrps         = Protein POPC SOL ION
> nstcalcenergy            = 1
> nstlist                  = 1
> nstcomm                  = 1
> comm_mode                = Linear
> comm-grps                = Protein_POPC    Water_and_ions
> ns_type                  = grid
> rlist                    = 1.2
> rlistlong         = 1.4
> vdwtype             = Switch
> rvdw                     = 1.2
> rvdw_switch         = 0.8
> coulombtype              = pme
> rcoulomb                 = 1.2
> rcoulomb_switch         = 0.0
> fourierspacing         = 0.15
> pme_order         = 4
> DispCorr             = no
> tcoupl                   = nose-hoover
> nhchainlength            = 1
> tc-grps                  = Protein_POPC    Water_and_ions
> tau_t                    = 0.5            0.5
> ref_t                    = 310         310
> Pcoupl                   = parrinello-rahman
> Pcoupltype               = semiisotropic
> tau_p                    = 5.0
> compressibility          = 4.5e-5       4.5e-5
> ref_p                    = 1.0          1.0
> pbc             = xyz
> gen_vel                  = no
> optimize_fft         = no
> constraints              = hbonds
> constraint_algorithm     = Lincs
> *


This discussion should be started with you reporting the speedup for
your system with 200cpu's. Rather than appealing just to some velocity
per nanosecond.

Neighborsearching is inefficient. Outputting immediate velocities
every 1000000 steps is senseless.


Dr. Vitaly V. Chaban, 430 Hutchison Hall, Chem. Dept.
Univ. Rochester, Rochester, New York 14627-0216
THE UNITED STATES OF AMERICA